Combining Dictionaries and Contextual Information for Cross-Lingual Lexical Substitution

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 117–122, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics USPwlv and WLVusp: Combining Dictionaries and Contextual Information for Cross-Lingual Lexical Substitution Wilker Aziz University of S˜ao Paulo S˜ao Carlos, SP, Brazil wilker.aziz@usp.br Lucia Specia University of Wolverhampton Wolverhampton, UK l.specia@wlv.ac.uk Abstract We describe two systems participating in Semeval-2010’s Cross-Lingual Lexical Substitution task: USPwlv and WLVusp. Both systems are based on two main com- ponents: (i) a dictionary to provide a num- ber of possible translations for each source word, and (ii) a contextual model to select the best translation according to the con- text where the source word occurs. These components and the way they are inte- grated are different in the two systems: they exploit corpus-based and linguistic resources, and supervised and unsuper- vised learning methods. Among the 14 participants in the subtask to identify the best translation, our systems were ranked 2nd and 4th in terms of recall, 3rd and 4th in terms of precision. Both systems out- performed the baselines in all subtasks ac- cording to all metrics used. 1 Introduction The goal of the Cross-Lingual Lexical Substitu- tion task in Semeval-2010 (Mihalcea et al., 2010) is to find the best (best subtask) Spanish transla- tion or the 10-best (oot subtask) translations for 100 different English source words depending on their context of occurrence. Source words include nouns, adjectives, adverbs and verbs. 1, 000 oc- currences of such words are given along with a short context (a sentence). This task resembles that of Word Sense Dis- ambiguation (WSD) within Machine Translation (MT). A few approaches have recently been pro- posed using standard WSD features to learn mod- els using translations instead of senses (Specia et al., 2007; Carpuat and Wu, 2007; Chan and Ng, 2007). In such approaches, the global WSD score is added as a feature to statistical MT systems, along with additional features, to help the system on its choice for the best translation of a source word or phrase. We exploit contextual information in alternative ways to standard WSD features and supervised ap- proaches. Our two systems - USPwlv and WLV usp - use two main components: (i) a list of pos- sible translations for the source word regardless of its context; and (ii) a contextual model that ranks such translations for each occurrence of the source word given its context. While these components constitute the core of most WSD systems, the way they are created and integrated in our systems differs from standard ap- proaches. Our systems do not require a model to disambiguate / translate each particular source word, but instead use general models. We experi- mented with both corpus-based and standard dic- tionaries, and different learning methodologies to rank the candidate translations. Our main goal was to maximize the accuracy of the system in choos- ing the best translation. WLVusp is a very simple system based es- sentially on (i) a Statistical Machine Translation (SMT) system trained …

Content truncated to 500 words